Anthropic research introduces GRAM for isolating dangerous AI knowledge

2026-07-09 · Source: Dataconomy · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Advanced, quick

Summary

Anthropic and AE Studio introduced Gradient-Routed Auxiliary Modules (GRAM), a method for isolating dangerous knowledge within AI models into removable modules without affecting overall performance. GRAM integrates small auxiliary neural compartments into a language model's MLP layers, each focusing on specific sensitive categories like virology or cybersecurity. During training, only the module corresponding to dual-use data, which made up approximately 0.25% of training data for each domain, is active. Researchers tested GRAM on models from 50 million to 5 billion parameters, including an 800-million-parameter model. Results showed that removing GRAM modules eliminated specific capabilities nearly as effectively as if the model had never been trained on that data, while general performance remained close to baseline. GRAM proved robust against adversarial fine-tuning, offering a potential compromise for AI policy by enabling selective access control, though it is preliminary and not yet applied to production models, with scalability and entangled capability separation remaining challenges.

Key takeaway

For AI Governance Policy Makers and AI Security Engineers grappling with dual-use AI risks, Anthropic's GRAM method presents a novel approach to granularly control sensitive knowledge. You can consider advocating for or implementing systems that allow selective module removal, enabling tailored model deployments for specific contexts like vetted research labs. This could mitigate national security concerns more effectively than broad restrictions, but be aware of current scalability limitations for production models.

Key insights

GRAM isolates dangerous AI knowledge into removable modules, maintaining general performance and offering selective access control.

Principles

• Knowledge can be modularly isolated in AI models.
• Targeted unlearning can be robust against fine-tuning.
• Selective access control enhances AI governance.

Method

GRAM extends transformer MLP layers with auxiliary modules. During training, only the module corresponding to dual-use data is active, allowing for later activation or deletion.

In practice

• Deploy models with specific knowledge modules removed.
• Grant vetted entities access to sensitive modules.
• Test modular unlearning against adversarial attacks.

Topics

• Gradient-Routed Auxiliary Modules
• AI Safety
• Dual-Use AI
• Knowledge Isolation
• AI Governance
• Transformer Architecture

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Security Engineer, Policy Maker

Related on AIssential

• Jul 8, 2026AlignmentAn off switch for dual use knowledge in AI models — GRAM enables configurable AI models by localizing dual-use knowledge into removable modules, allowing selective capability control from a single training run.
• Cybersecurity Experts Furious After U.S. Bans Anthropic’s Fable — Banning defensive AI tools based on misinterpreted "jailbreaks" undermines cybersecurity without preventing misuse.
• I didn’t expect this from Anthropic — Anthropic's internal data shows AI is accelerating its own development, raising concerns about recursive self-improvement and the need for a coordinated pause.
• Anthropic’s Safety Superpower — Anthropic's "safety" narrative strategically aligns with its economic and power imperatives to control frontier AI development and user data.
• Anthropic Added a New Security Measure to Get Back Into the Trump Administration’s Good Graces — AI model developers face increasing regulatory scrutiny requiring robust, verifiable security measures to maintain operational access.
• Saudi firm closes first fund since start of Iran war — Anthropic's Fable 5 demonstrates that powerful AI models can be safely deployed for general use through robust guardrails and tiered access.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Dataconomy.